Improve how our HTTP proxy code reads the content-length header.

It now reads the length into an unsigned variable. This was reported
to us by Matt Jones of Volvent. I think there's no potential for
remote code execution or remote crash--we don't allocate a buffer
for the content or anything. I think it's just ugly code. I didn't
impose a max size for the content-length. I feel like it's ok for the
while loop to loop many times then get EAGAIN and break out.

I'm looking forward to unified http parsing in 3.0.0.